This study presents our team (FuneLab) submission to the 2026 George B. Moody PhysioNet Challenge. Overnight polysomnography captures changes in sleep, respiration, cardiovascular activity, and movement that may be associated with future cognitive impairment. We developed a multi-rate, channel-wise 1-D convolutional network that processes raw EEG, EOG, EMG, ECG, respiratory, and SpO2 signals at sampling rates suited to their physiological dynamics. Learned representations are fused progressively as their temporal resolutions become compatible. The model also incorporates demographic information, algorithmic sleep annotations, and SpO2 features describing baseline and desaturation burden. Explicit availability indicators represent missing channels and annotations. Predictions from multiple two minute windows are aggregated to produce a patient-level risk estimate. Our highest scoring official submission on the primary metric was trained on the small dataset and achieved an age-conditioned AUROC of 0.729 and a Reward of 0.253 on the hidden validation set. These results demonstrate the feasibility of combining raw PSG waveforms with physiological and sleep related data for cognitive impairment prediction.